Page 303 - Vibrational Spectroscopic Imaging for Biomedical Applications

Raman Imaging for Biomedical Applications in Clinics 277

on spectral data analysis methods the reader is referred to existing litera-
ture, since that is beyond the scope of this chapter. 31

9.4.1 Classification Techniques
The currently most used technique to generate Raman images is clus-
ter analysis. Cluster analysis is an unsupervised classification tech-
nique that identifies groups of spectra, based on their spectral
similarity. Group membership can be color coded, giving similar
spectra the same color. Different tissue structures and tissue patholo-
gies will have a different biochemical makeup, which is reflected in
their Raman spectra, and therefore the similarity of spectra within
a certain structure will be greater than the similarity between spec-
tra of two different structures. The resulting Raman image there-
fore shows tissue structures and pathologies as a pseudocolor
image. The spectral information can further be used to analyze the
difference between the mean spectra of different structures and
pathologies and thus provide chemical information about these
differences. Note that no spatial information is used in this type of
analysis.
Cluster analysis results are dependent on both the clustering
algorithm used, and on the similarity (or distance) measure used.
There are two types of cluster algorithms: hierarchical and nonhierar-
chical, partitioning algorithms. Hierarchical algorithms like single
linkage, complete linkage and Ward’s method, calculate the similar-
ity or distance between any two of the spectra in the Raman map (full
similarity or distance matrix). For large maps this becomes a compu-
tational intensive procedure. Partitioning methods like k-means and
ISODATA are less computationally demanding, because they do not
require a full distance matrix. These algorithms assign all spectra of a
map to the closest of a number of predefined spectral cluster centers
(which, for instance, can be spectra from a number of different loca-
tions in the map). Then they iterate through all data points, updating
the cluster centers as spectra are assigned to the clusters, until a stable
solution is reached. Partitioning methods are generally faster for big
Raman maps, but they are heuristic (result is dependent on the choice
for the initial cluster centers) and not complete (the number of clus-
ters has to be predefined, whereas hierarchical cluster results in a
complete membership matrix, and the desired amount of clusters can
be determined afterward).
If one wants to classify the spectra of an image into predefined
32
groups, techniques like linear discriminant analysis and artificial
33
neural networks can be used. These are supervised techniques
that make use of a multivariate model based on a gold standard to
classify the spectra that are used for building that model. The
model then extracts the spectral features that are most useful to
discriminate between these different groups, and this information can
then be used to classify new spectra into one of the groups. Using these